GDPR-compliant AI chatbot: a UK checklist

An AI chatbot that talks to your customers processes personal data from the first message. Names, email addresses, order numbers and complaints all count. That does not make AI support risky, but it does mean the same UK GDPR rules apply as for any other system that handles customer data.

This checklist covers what to sort out before you launch, and the questions to put to any vendor. It is practical guidance, not legal advice; if you are unsure, speak to your data protection officer or a solicitor, and read the ICO's guidance on AI and data protection.

1. Know what data the chatbot will see

Map the data before you choose the tool. Typical categories are:

• Contact details the customer volunteers or you pass in from a logged-in session
• Order, booking or account references and the data returned when the bot looks them up
• The free text of the conversation, which can contain anything, including special category data such as health information
• Technical data such as IP address, browser and page visited

If customers might share health, financial hardship or similar sensitive details, plan for it: tell them not to, and make sure the agent hands those conversations to a person.

2. Pick a lawful basis

For customer support, legitimate interests or performance of a contract usually fit: the customer is asking you for help with something you sell them. Record your reasoning. If you use conversation data for anything beyond answering the customer, such as marketing, you will probably need a separate basis and, for marketing, consent.

3. Put a processor contract in place

Your chatbot vendor is almost certainly a processor acting on your instructions. Article 28 of the UK GDPR requires a written contract, usually a data processing agreement (DPA), covering what they process, security, sub-processors, deletion at the end of the contract and help with rights requests. Ask for it before you go live, not after.

Also ask for the sub-processor list. AI agents typically rely on a model provider such as OpenAI, Anthropic or Google, plus hosting and email providers.

4. Check where the data goes

If personal data leaves the UK, you need a lawful transfer route: an adequacy decision, or safeguards such as the UK International Data Transfer Agreement or the UK Addendum to the EU standard contractual clauses. Ask the vendor:

• Where conversation data and uploaded documents are stored
• Where model requests are processed, and under what terms
• Whether model providers retain prompts or use them for training

A vendor that hosts in the UK or EU and uses model providers under zero-retention API terms makes this part much simpler.

5. Be transparent

Customers should know they are talking to an AI and how to reach a person. Update your privacy notice to cover the chatbot: what you collect, why, who processes it and how long you keep it. A short line in the widget itself, linking to the notice, is good practice.

6. Keep a human in the loop for decisions that matter

Article 22 restricts decisions based solely on automated processing that have legal or similarly significant effects. A bot answering "where is my parcel?" is fine. A bot refusing a refund, cancelling an account or assessing a claim on its own is a different matter. Design escalation rules so those conversations reach a person, and log when they do.

7. Minimise and set retention

Only pass the chatbot the data it needs. If an action looks up an order, return the status and tracking link, not the customer's full profile. Decide how long you keep transcripts, and make sure deleting a bot, a source or an account actually deletes the underlying data, including embeddings.

8. Be ready for rights requests

Customers can ask for a copy of their data or for it to be erased. You should be able to find a person's conversations, export them and delete them without a support ticket to your vendor.

9. Consider a DPIA

A data protection impact assessment is required where processing is likely to result in a high risk. The ICO lists innovative technology, including AI, as a factor to consider. Even where it is not strictly required, a short DPIA is a good way to document the decisions above.

10. Secure it

• Encryption in transit and at rest
• Role-based access for your team, with viewer roles for people who only need to read
• Signed webhooks for anything that calls your systems
• Guardrails against prompt injection and system prompt leaks

Questions to ask any AI chatbot vendor

1. Will you sign a DPA, and can I see your sub-processor list?
2. Where is my data hosted, and where are model requests processed?
3. Do your model providers retain or train on my data?
4. How do I export or erase one customer's conversations?
5. What happens to my data when I delete a bot or close my account?
6. How does the agent hand sensitive conversations to a person?

How Threadbox handles this

Threadbox is hosted in the UK/EU, encrypts data in transit and at rest, and uses model providers under zero-retention API terms. Your content is used only to answer your customers. A DPA is available on request, conversations can be exported or erased per visitor, and deleting a bot removes its sources, embeddings and conversations. Escalation rules route sensitive or high-stakes conversations to your team. Read more on the security page.